Study design

This study was approved by the THIN Scientific Review Committee on the 16th of June 2014 (Reference Number 14–043). A cohort study was designed on routinely collected data of primary care patient records from The Health Improvement Network (THIN) UK database (http://www.thin-uk.net/). The database currently holds 12 million patient records from 578 general practices, corresponding to 6% of the UK population, with slightly more THIN patients living in affluent areas compared to nationally [16]. THIN is generalizable to the UK population with regards to demographics, prevalence of medical conditions, and mortality rates, adjusted for deprivation [17–18].

Medical records from 1987 to 2011 of people who were born between 1920 and 1940 were selected. Four age cohorts were created that consisted of participants who turned the target age between 1^st January 1987 and 2011, and participants could be part of multiple age cohorts. The target ages were 60, 65, 70, and 75 in order to facilitate individuals and general practitioners to make a decision about statin uptake at key ages. Before age 60, the initiation rate of statin prescription for primary prevention of cardiovascular disease in the UK is less than 3% [19].

Participants were selected if at the target age they had been registered for at least one year at an active general practice that coded death dates validly, and their medical records had been accessed at least once within the last ten years. Participants with a previous history of cardiovascular disease (coronary artery disease, cerebrovascular disease, or peripheral vascular disease) were excluded from the analysis, as were participants living in Northern Ireland and Scotland due to unavailable Townsend deprivation scores. Participant selection was performed in SQL server 2012. Participants were followed up until the 18^th of March 2011, resulting in a study period of 24, 24, 21, and 16 years for the 60-, 65-, 70-, and 75-year old cohorts, respectively. Participants were lost to follow-up when they transferred to another general practice during the observation time. It was assumed that this was independent from survival time.

Risk assessment of cardiovascular disease

The NICE guideline on lipid modification recommends using QRISK2 to calculate the 10-year risk of a first cardiovascular event [6]. The QRISK2 algorithm is available online (http://svn.clinrisk.co.uk/opensource/qrisk2/). It incorporates information on multiple demographic, medical, and lifestyle factors to estimate the risk (S1 Table) [20]. Not all this information was available or complete in THIN, leading to a number of substitutions or exclusions. The QRISK2 was calculated for all participants with complete information in JAVA version 8.

Ethnicity was not recorded in THIN and could therefore not be included in the QRISK2 calculation. The QRISK2 calculator has white ethnicity as the reference category. 93% of the UK population is white and the QRISK2 score risk would be underestimated for people with an Indian, Pakistani, or Bangladeshi background and overestimated for people with a black Caribbean background and for black African and Chinese men [20]. In THIN the Townsend score was available in quintiles based on the 2001 census data, and the associated median value for each quintile was used for calculating the cardiovascular risk [21]. The prevalence of atrial fibrillation and rheumatoid arthritis is <1% in the United Kingdom, making their effect on QRISK2 minimal at population level [20,22], and it was assumed that the selected participants did not have these medical conditions. Family history of cardiovascular disease is not systematically recorded in primary care, and was <1% in the four cohorts in THIN [20]. It was assumed that diabetes was type two, which would be true for 90% of the cases [23]. It was also assumed that current smokers smoked moderately [24]. The diagnosis of hypercholesterolaemia was used instead of a specific cholesterol level, because it was only recorded in the minority of primary care records (27–50% recordings) [20]. When a participant did not have a diagnosis of hypercholesterolaemia, a ratio of total cholesterol to high density lipids of four was ascribed. This is the default value of the online QRISK2 calculator when no ratio is given. When a participant had a diagnosis of hypercholesterolaemia, a ratio value of five was ascribed.

Statistical analyses

Cox’s proportional hazard regression was used to estimate the effect of statin prescription on the hazard of mortality for different risk groups at various ages, with the outcome variable being time to death in days. All-cause mortality was used because other studies have shown the protective effect of statins on the risk of cardiovascular and cerebrovascular events and deaths, but it was still unclear what the overall survival effect is. All-cause mortality is not only an important outcome for patients, but also for the national economy. Life expectancy directly affects taxation, national insurance rates, retirement planning, including state and private pensions, and actuarial products such as life insurance and annuities. Furthermore, death is an unambiguous hard outcome that is not subject to medical classifications such as coronary revascularisation procedures, that depend on rates of diagnosis that are subject to differing professional opinions and geographic variations [25]. The exposure was any statin prescription at any time before the participant reached the baseline age. The analysis on statins excluded participants who were prescribed other lipid-lowering therapy to ensure that the effect of statins was not mediated by those drugs. The number of participants excluded at age 60, 65, 70, and 75 were 876 (0.7%), 1,718 (0.9%), 2,840 (1.1%), and 2,475 (1.3%), respectively. The analysis performed was on an intention to treat basis as it was based on the prescription of drugs, and not intake of them. The models were fitted to each age and risk group. The ages were 60, 65, 70, and 75 years. The risk groups consisted of participants with a QRISK2 of <10%, 10–19%, or ≥20% risk of a first cardiovascular event in the next ten years.

Potential confounders of the effect of statin prescription on survival were selected on the basis of expert knowledge and literature review (S2 and S3 Tables). With backward elimination, the variables that were significant (p<0.05) in the majority of the models, including interactions with statin prescription, were adjusted for in the final models. The final models adjusted for sex, year of birth, socioeconomic status, diabetes, hypercholesterolaemia, blood pressure regulating drugs, body mass index, smoking status, and general practice. Socioeconomic status was measured by Mosaic, a classification that captures demographics, lifestyles and behaviour of people at postcode level [26]. In the preliminary results, Mosaic was found to be a stronger predictor of mortality than Index of Multiple Deprivation (IMD) and Townsend deprivation score. There were missing values in systolic blood pressure (15–28%), smoking status (14–30%), and body mass index (22–37%). The fraction of incomplete medical records decreased with age; 46% of the youngest cohort and 29% of the oldest cohort had incomplete records. The proportion of complete records was greater among participants born at a later year, with a medical condition, or who were prescribed drugs (S4 Table). This is in agreement with previous research that found that recording is likely be related to ill health and that recording improved after the introduction of the Quality and Outcomes Framework (QOF) in 2004 [27–29]. Incomplete medical records were dealt with by multiple imputation (S5 Table) [30]. The distribution of known and imputed values for systolic blood pressure, body mass index, smoking status, and QRISK2 scores were similar (S6 Table).

The survival models were multilevel on general practice and participant level, to deal with the interdependence of participants from the same general practice [30]. Efron’s approximation was used to deal with time-tied deaths [31]. The number of years gained or lost due to statin prescription were calculated [32]. The models were checked for proportional hazards assumption, overall performance, discrimination, and external validation [33–36]. The sensitivity analysis included comparing the adjusted effects estimated on the complete and incomplete records with (1) the unadjusted effects estimated on the complete and incomplete records, and (2) the adjusted effects estimated on the complete records. The analysis was executed in R version 3.1.1 using the packages ‘survival’ [31] and ‘rms’ [36].

For more information on the multiple imputation, model assumptions, and model assessments, please see S5 Table.

Strengths and limitations

This study used routinely collected primary care data that were representative of the UK population and widely available. The large sample size included a great number of participants aged over 80 and many people at low risk of a cardiovascular event, with almost 25 years of follow-up data. The 10-year risk of a first cardiovascular event was calculated by QRISK2, which is recommended by NICE and widely used in routine practice, and broadly comparable to other widely used risk assessment tools such as SCORE and PCE. The use of all-cause mortality, meant that the overall effect of statin prescription on mortality could be assessed as opposed to estimating the possible shift of the hazard of mortality from one medical condition to another. Estimating the effect of statin prescription by age group meant that age-group specific recommendations could be given.

The analysis was performed on an intention to treat basis to more accurately assess the effect of routine current practice in the general population. THIN had information on prescription of drugs, and not on dispensing and intake of them. This means that the actual statin uptake could be lower than THIN records indicated. The analysis did not include the duration of therapy as a possible predictor of survival. As the majority of the patients had started statin therapy within the six years prior to each key age for each risk group, it seems more likely that the benefit of statins is age-dependant rather than duration-dependant. In other words, it is unlikely that the younger participants did not experience a survival benefit from statins due to a shorter duration of therapy. Due to limitations of the data, the QRISK2 score could only be approximated. Missing data are described in the methods section and S5 Table. Although there were missing data, sensitivity analyses showed that it was unlikely they influenced the results.

Limitations of a prevalence-user (rather than new-user) study design are that bias might be introduced when the exposure’s effect on the outcome varies by time and alters other health indicators [38]. As the hazard of mortality associated with statin prescription did not differ by time, it is unlikely that underascertainment of deaths between the time statin therapy is initiated and the target age is reached would have biased the results.

Randomised controlled trials are rightly considered to be a stronger form of evidence than cohort studies, primarily because randomisation is designed to ensure that potential confounders are evenly distributed between control and intervention groups. A limitation of using routinely collected observational data to estimate the effects of interventions is that the results might be affected by unexplained confounding. This was minimised by stratifying by cardiovascular risk group and age, and by controlling in the regression models for a wide range of potential confounders. Despite this, we cannot completely exclude residual confounding by indication. As this is an observational study, both sick-user bias and healthy-user bias cannot be excluded. Sick-user bias may arise where at a given QRISK2 score, general practitioners recognised people who were at greater risk and consequently prescribed them statins. This bias would have led to an underestimation of the survival benefit of statin prescription. Healthy-user bias would arise where at a given QRISK2 score, patients who were more proactive about their health, were more likely to be treated with a statin [38–39]. This bias would have led to an overestimation of the survival benefit of statin prescription. We believe that both biases were mitigated by inclusion of co-morbidities, lifestyle factors and socioeconomic status in our models but cannot be excluded.

These findings reinforce the considerable benefits of statin treatment in high-risk groups (where we found substantial and important undertreatment) reported by the extremely influential CTT meta-analysis [5], but clearly differ from the CTT results for those at low risk. Where we found no benefit, despite widespread treatment of participants (particularly women) at low risk, the CTT reported an overall reduction in all-cause mortality (rate ratio 0.91) in participants without a history of cardiovascular disease over 2–5 years, and concluded that statins ‘are effective for people with a 5-year risk of major vascular events lower than 10%’ [5]. A more recent review reiterated the well-known methodological limitations of observational studies of treatment [40], which we agree with and have discussed above. However, the use of statins in primary prevention remains controversial and contested, perhaps at least in part due to the limitations of RCTs discussed in the introduction, including lack of generalisability due to strict inclusion criteria, lack of comparability with clinical risk scores such as QRISK2 due to trials use of observed events as a comparator (the CTT trial 5-year risk of major vascular events lower than 10% is hard to compare to a general population QRISK2 score of 10% and is not equivalent), the small number of older patients in trials, and the relatively short follow-up time of 4–5 years in trial (compared with 6–12 years in the age cohorts studied here), and perhaps most importantly, the concerns that anonymised individual patient data from statins trials have still not been made available for independent scrutiny, and remain under the control of a single group of respected researchers, whilst statins are among the most widely prescribed drugs globally [7, 41]. Large scale observational studies of the effects of statins in everyday diverse clinical practice over many years are an under-explored source of information on the effects of statins on mortality. No data source is perfect and there are well rehearsed uncertainties and unanswered questions arising from both observational and trial data (set out in the introduction), and we believe that this analysis of cohort data fills in some of the gaps and provides an important new source of information on the possible effects of statins in routine practice.

Recommendations

Further research is needed on the effects of statins over the long term for younger patients at low risk of a future first cardiovascular event. These findings suggest that the current internationally recommended thresholds for statin therapy for primary prevention of cardiovascular disease in routine practice are too low, and may lead to overtreatment, particularly of people under 60 and at low (<10%) risk. The recent revision of guidelines to extend treatment to younger and lower risk groups may need to be reconsidered, and clinicians may want to use this new information when discussing the risks and benefits of statin initiation with their patients.